Report: Anthropic Releases Research Preview of ‘MHS’ Device-Connection Standard for AI Agents

tmtpost.com reported on September 4, 2026, that Anthropic had released a research preview of “MHS,” a device-connection standard for AI agents. We have not been able to verify the original source, so please refer to it for details.

In the reported example, individual connection code was needed to link experimental equipment that used different control methods. MHS is said to be designed to let agents handle equipment capabilities and constraints in a readable format. For engineers integrating equipment in laboratories and manufacturing sites, a specific issue is how to standardize connection work and the description of safety constraints.

The reported key feature of MHS is that it places a common interface for AI agents to operate physical equipment on top of equipment-specific control systems. The article describes a design that organizes equipment status and operations into read and write commands, and presents what can be measured, which parameters can be changed and what safety constraints apply in a format agents can access. MHS itself, however, is said to be neither a robot operating system nor a real-time joint-control system. The technical feature highlighted in the report is the division between the scope handled by the connection standard and that handled by existing control software. This division of labor differs from a concept in which a connection standard alone replaces the entire operation of robots or experimental equipment. In the reported example, the agent plans equipment operations, while deterministic code and existing control systems handle high-speed execution and long-duration operation. Therefore, the evaluation is likely to focus not only on the number of devices that can be connected, but also on how accurately equipment descriptions represent actual capabilities and safety constraints, and whether device state is preserved when control is handed off. The article does not, however, cover all of these operating conditions or the scope of validation. In a later article, we reported on an example of applying Anthropic’s system to robot operation in QuEra demonstrates Anthropic’s “physical AI version of MCP,” achieving a 99% success rate in error recovery. This is a separate system and example and does not indicate technical equivalence with MHS. We also covered plans to combine chemists, robots and AI for drug discovery research in Warp Sciences to build a robotics-enabled drug discovery research platform, raises $215 million. The connection between the two articles is that linking experimental equipment with AI is relevant to research workflows, but no relationship, such as adoption of MHS in the company’s plans, has been confirmed. An implication for the industry is that converting the different control methods and documentation used by equipment makers into a common description that agents can handle could become a new step in deployment. The article said that in a system at Carnegie Mellon University, each device had a different connection method and required individual integration code. Standardization could reduce custom integration work, but a single reported example is not enough to determine how widely the approach can be reused across devices from different manufacturers. After MHS’s release, points to check include supported equipment and implementation examples, who is responsible for maintaining the descriptions, methods for validating safeguards against incorrect operation, and the division of responsibility when experimental data is used to inform subsequent decisions. The article’s reported test duration included a comparison with the several weeks suppliers would take to build a similar system, but details of the comparison conditions were not provided. Please refer to the original source for details.